Milky Way structure simulations earn Gordon Bell nomination
If you took a photograph of the Milky Way galaxy today from a distance, the photo would show a spiral galaxy with a bright, central bar (sometimes called a bulge) of dense star populations. The Sun—very difficult to see in your photo—would be located outside this bar near one of the spiral arms composed of stars and interstellar dust. Beyond the visible galaxy would be a dark matter halo—invisible to your camera but important nonetheless because it keeps everything together by dragging down the rotational velocity of the bar and spiral arms.
Now, if you wanted to go back in time and take a video of the Milky Way forming, you could go back 10 billion years, but many of the galaxy’s prominent features would not be recognizable. You would have to wait about 5 billion years to witness the formation of the Earth’s solar system. By this point, 4.6 billion years ago, the galaxy looks almost like it does today.
“The grand structure of the galaxy has emerged from the self-organization of the stellar distribution over the past 10 billion years to eventually look like the Milky Way in the photo,” said Simon Portegies Zwart of Leiden Observatory in the Netherlands.
This is the timeline a team of researchers from the Netherlands and Japan, including Portegies Zwart, are seeing emerge when they use supercomputers to simulate the Milky Way galaxy’s evolution. Using a code developed for GPU supercomputing architectures—including that of the Cray XK7 Titan located at the Department of Energy’s Oak Ridge National Laboratory—the team’s simulations have earned acceptance as a Gordon Bell Prize finalist. The prize recognizes outstanding achievement in high-performance computing and will be presented by the Association for Computing Machinery at SC14 on November 20.
“We don’t really know how the structure of the galaxy came about,” Portegies Zwart said. “What we realized is we can use the positions, velocities, and masses of stars in three-dimensional space to allow the structure to emerge out of the self-gravity of the system.”
The challenge of computing galactic structure on a star-by-star basis is, as you might imagine, the sheer number of stars in the Milky Way—at least 100 billion. Therefore, the team needed at least a 100 billion-particle simulation to connect all the dots. Before the development of the team’s code, known as Bonsai, the largest galaxy simulation topped out around 100 million—not billion—particles.
The team tested an early version of Bonsai on the Oak Ridge Leadership Computing Facility’s Titan, the second-most-powerful supercomputer in the world, to improve scalability in the code. After scaling Bonsai to almost half of Titan’s GPU nodes, the team ran Bonsai on the Piz Daint supercomputer at the Swiss National Supercomputing Centre and simulated galaxy formation over 6 billion years with 51 million particles representing the forces of stars and dark matter. After a successful Piz Daint run, the team returned to Titan to maximize the code’s parallelism.
The Bonsai code demonstrated scalability on 18,600 Titan nodes (96% of the machine’s GPU nodes), which would enable an 8 million-year, 242 billion-particle Milky Way simulation. Bonsai achieved nearly 25 petaflops of sustained single-precision, floating point performance on Titan. Single-precision floating-point operations use less memory by representing numbers using 32 bits, whereas double-precision operations represent more precise numbers at the expense of using 64 bits.
“With graduate student Jeroen Bédorf, we started by writing single code for GPUs and deliberately never wrote code on CPUs because we wanted the entire code to run on GPUs to exploit their parallelism,” Portegies Zwart said. “The host CPUs are only used to streamline the communication between the nodes and the GPUs. In this way, we can fully optimize the use of GPUs for number crunching and the much slower CPUs to minimize the communication overhead.”
Another feature of Bonsai that makes 242 billion-particle simulations feasible is the use of a hierarchal tree-code that eliminates direct gravitational force calculations between each particle (or star) and all the other billions of particles by organizing them into octants that prioritize particle interactions.
The team aims to compare simulation results to new observations coming from the European Space Agency’s Gaia satellite that launched last year. The Gaia mission is currently cataloguing star measurements—including distances, velocities, and stellar type—of one billion Milky Way stars.
“One percent of the particles, or stars, in our simulated galaxy should match Gaia data,” Portegies Zwart said.
Gaia will also provide data on stars farther than Earth’s Solar Neighborhood, or only stars within tens of light-years. When compared with Bonsai simulations, these new observations will help researchers better understand larger galaxy dynamics, such as the interaction of the bar and spiral arms, in addition to local dynamics taking place around the Solar Neighborhood.
Because the galaxy is a big place with many research teams dedicated to understanding it, Portegies Zwart’s team plans to make simulation data and source code resulting from Bonsai projects available to the research community.
The research team that developed the application nominated for the Gordon Bell Prize includes Portegies Zwart and Bédorf of Leiden Observatory, Evghenii Gaburov of SURFsara Amsterdam, Michiko S. Fujii of the National Astronomical Observatory of Japan, Keigo Nitadori of RIKEN Advanced Institute for Computational Science, and Tomoaki Ishiyama of the University of Tsukuba.
This work was supported by the Netherlands Research Council; the Netherlands Research School for Astronomy; the National Astronomical Observatory of Japan; and Japan’s Ministry of Education, Culture, Sports, Science and Technology. This research made use of the Oak Ridge Leadership Computing Facility, a Department of Energy Office of Science User Facility at Oak Ridge National Laboratory.
UT-Battelle manages ORNL for the DOE Office of Science. The Office of Science is the single largest supporter of basic research in the physical sciences in the United States, and is working to address some of the most pressing challenges of our time. For more information, please visit https://science.energy.gov/.